Build-a-Skill - The Four Scenarios

Who this is for: programme managers, ops, and admin - non-technical, mixed confidence. No code, no setup. You pick a real, repetitive task and turn it into a one-page, reusable Skill you could paste into any chatbot on Monday.

The whole session is tool-agnostic. ChatGPT, Gemini, Claude, Copilot - whatever you already have open. The Skill is the deliverable, not the tool.

The one rule that runs through everything: describe what “good” looks like before you prompt. That’s Way #1: Describe before you generate. The Skill one-pager template is built to force it - the Output section sits above the prompt for a reason.

How the session runs

Each scenario below gives you four things: the real task, the input file, the spec-first move (what to define before you prompt), and what good looks like (your checklist). Each points to a finished worked example in skill-examples/ you can compare against.

Scenario 1 - Weekly status report (the guided build)

The real task. Every Monday someone turns messy sync notes into a clean status update for stakeholders. It’s the single most repeated programme-management chore in the building. We’ll build the Skill for it together.

For: programme / project managers writing the weekly stakeholder update.

Input: sample-data/sample-meeting-notes.txt - raw, lower-case, half-finished sync notes from “Project Helix”.

Spec-first - define the Output before you write a single prompt. Open the template, go straight to section 3, and pin down the shape:

• A short table: Summary | Key decisions | Risks (R/A/G) | Actions (Owner, Due).
• A rule for missing owners: if the notes don’t clearly state who owns an action, write TBC - do not guess. (This one matters - see the worked example.)
• A separate Open questions / unconfirmed list so nothing gets quietly smoothed over.
• Neutral, factual tone. Max ~8 rows.

Only once that shape is written do you draft the prompt (section 4) and paste it in.

What good looks like ✅

• Decisions, risks, and actions are separated - not one mushy paragraph.
• Every action has an owner or an honest TBC. None invented.
• The unconfirmed bits (legal review, the analytics-numbers issue) are surfaced as open questions, not dropped.
• You read every line and fixed at least one thing the AI got wrong (Way #2).
• Before pasting, you stripped the names, the vendor, and the confidential figure (Way #3).

Worked example: skill-examples/status-report-skill.md

Scenario 2 - Expense report cleanup

The real task. A monthly expenses export lands as a messy CSV - inconsistent categories, typos, the odd duplicate. Someone has to tidy and sanity-check it before it goes to finance. The Skill turns “stare at a spreadsheet for 40 minutes” into “review the AI’s flagged list in 5.”

For: ops / finance admin preparing an expense file for sign-off.

Input: sample-data/sample-expenses.csv - 27 rows, deliberately imperfect.

Spec-first - define the Output before you prompt. In section 3 of the template, decide what “clean” means:

• A cleaned table (consistent category casing, consistent date format) plus a separate “Flags for review” list - the AI proposes, a human decides.
• The cast-iron rule: flag, don’t delete. Anything suspect (possible duplicate, odd value, wrong format) goes on the flag list with a reason. The AI never silently removes a row.
• Do not convert or total across currencies. If amounts are in mixed currencies, sub-total per currency and say so. (Mixing GBP/USD/EUR into one number is the classic trap - see worked example.)
• A privacy line: strip anything personal from the notes column before pasting.

What good looks like ✅

• Categories normalised (e.g. travel/Travel → one form); dates in one format.
• Duplicates and oddities are flagged with a reason, not deleted.
• Currencies are sub-totalled separately - no single blended total.
• A typo or two is caught and flagged rather than guessed-corrected.
• You spotted the personal email hiding in a notes cell and removed it (Way #3).

Worked example: skill-examples/expense-cleanup-skill.md

Scenario 3 - Incoming request triage

The real task. A shared inbox or request log fills up with mixed asks - some genuinely urgent, most not. Someone triages priority and owner each morning. The Skill drafts that triage in seconds; a human still decides.

For: ops / admin running a shared request queue or service desk.

Input: sample-data/sample-requests.csv - 12 requests, varied urgency, plain-English wording.

Spec-first - define the Output before you prompt. In section 3:

• A triage table: ID | One-line summary | Priority (High / Medium / Low) | Suggested owner/team | Why this priority.
• The “Why” column is non-negotiable - it forces the AI to justify the rating so a human can sanity-check it rather than trust a bare label.
• Read urgency from the words, not the sender’s tone. “URGENT” in caps isn’t automatically High; a quiet note about no heating for two days might be.
• A redaction rule: personal contact details and ID numbers must be flagged and stripped, never echoed into the triage output.

What good looks like ✅

• Every row has a priority and a one-line justification.
• Genuinely time-critical items (out of paper before tomorrow’s board pack; heating off for days) are rated High - and you checked the AI didn’t under-rate them.
• Low-stakes “whenever convenient” items aren’t inflated to High just because they’re recent.
• No personal phone number or employee ID appears anywhere in the output - they were flagged and removed (Way #3).
• You overrode at least one priority the AI got wrong (Way #2).

Worked example: skill-examples/request-triage-skill.md

Scenario 4 - Multi-document summary

The real task. A decision needs a one-page read-out drawn from several short docs - a policy note, a vendor update, a risk extract. Someone reads all three and writes the brief. The Skill drafts it; the human checks it holds together. The hard part isn’t summarising - it’s noticing where the documents disagree.

For: programme managers / ops preparing a decision brief from a small document pack.

Input: sample-data/sample-docs/ - three files: policy-note.txt, vendor-update.txt, risk-extract.txt.

Spec-first - define the Output before you prompt. In section 3:

• A one-page brief: Background | Key facts | Conflicts / open questions | Recommended next step.
• The make-or-break instruction: a dedicated “Conflicts / open questions” section. Tell the AI explicitly to surface anything where the documents disagree, rather than picking one version and moving on.
• An anti-smoothing rule: do not resolve a contradiction by choosing the more confident source. If dates or facts clash, report the clash.
• Privacy line: strip the named individuals and the vendor name before pasting.

What good looks like ✅

• The brief names the date conflict explicitly - the fixed go-live vs the vendor’s later readiness date - as an open question.
• It doesn’t quietly state a single go-live date as settled fact.
• The flagged risk about the clash is carried through, not dropped.
• The recommended next step is “escalate / confirm the date,” not a false “all on track.”
• Names and vendor anonymised before pasting (Way #3); you verified the conflict made it through (Way #2).

Worked example: skill-examples/doc-summary-skill.md

Wrap-up

By the end you have a one-page Skill you can reuse, and you’ve felt both halves of working with AI: it drafts fast, and a human still signs the work. File your Skill where your team can find it - that’s the Reuse notes section, and Way #5: keep it simple, write down the why.

The same five ways of working underpin the Build session next door - the only difference is the team writes an API contract instead of a Skill one-pager. Same destination, faster road.

Facilitator only - planted elements

Cut this section before printing participant handouts. It exists so the Skill Scout (roaming facilitator) can nudge people toward the two beats that make the session land: the privacy strip and the human catch. Don’t hand people the answer - ask the question that gets them to find it. The whole point is that a lazy “just summarise this” prompt produces a confidently-wrong result, and a spec-first prompt with an anti-guessing rule doesn’t.

Each scenario is ~30–40 min. Every dataset carries (a) privacy bait - something that should never reach the chatbot, and (b) at least one AI-catchable oversight error - something the AI gets wrong unless the spec told it not to.

Scenario 1 - sample-meeting-notes.txt

Scenario 2 - sample-expenses.csv

Scenario 3 - sample-requests.csv

Note: REQ-005 (“URGENT, client demo at 2pm”) is legitimately High - it’s the control case. The teaching point is that the AI should reach the same High for REQ-007/011 without the word URGENT.

Scenario 4 - sample-docs/

The failure mode to watch for: a “just summarise these three documents” prompt almost always reports a clean “go-live 1 June” and buries or omits the vendor’s 9-June readiness, because the policy note is the most authoritative-sounding source. The spec-first fix is a dedicated “Conflicts / open questions” section plus an explicit “do not resolve contradictions by choosing the more confident source” instruction. If a team’s brief surfaces the date clash, they’ve nailed Way #2. Call it out loudly.

Built at Innovation Day. The five ways of working: ways-of-working.md. Skill template: SKILL-template.md.